<p>Delayed recognition of morphometric abnormalities in breast cytology can contribute to disease progression and adverse clinical outcomes. Conventional cytopathological assessment is affected by subjectivity and inter observer variability, which highlights the need for automated and objective diagnostic support. This study proposes an explainable machine learning framework designed to analyze fine needle aspiration (FNA) cytomorphometric data in order to classify breast lesions and stratify malignancy risk. The Breast Cancer dataset was reorganized into four clinically relevant categories: Typical Benign, Atypical Benign, Early Malignant, and Advanced Malignant. This classification was defined based on hierarchical clustering applied to 30 nuclear and structural morphometric features. Class imbalance was addressed using SMOTE, Gaussian noise augmentation, and Mixup. Five lightweight classifiers were evaluated: LightGBM, XGBoost, multilayer perceptron, random forest, and AdaBoost. Hyperparameters were optimized through grid search with nested cross validation. Model performance was assessed using accuracy, precision, recall, F1 score, and area under the ROC curve (AUC), evaluated both on an independent test set and through stratified five fold cross validation. Statistical significance in AUC differences was examined using the DeLong test. Interpretability was ensured through SHAP, which provided global and instance level explanatory insights. LightGBM presented the best overall performance, achieving a mean accuracy of 97.07% (±0.62), an F1 score of 95.41% (±1.24), and an AUC of 0.987 (±0.004). Its advantage over XGBoost was statistically significant, with an AUC difference of 0.024 and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p = 0.017\)</EquationSource> </InlineEquation>. SHAP analysis indicated that concavity_mean, area_worst, and perimeter_worst were the most influential morphometric features associated with malignant lesions. The proposed approach introduces a transparent and efficient clinical decision support system for automated triage of breast lesions based on FNA derived morphometric data. The model enhances diagnostic precision and enables malignancy risk stratification, aligned with the principles of Predictive, Preventive and Personalized Medicine (PPPM/3PM). Its explainable design, computational efficiency, and compatibility with telemedicine and digital cytology workflows support its integration into routine practice and favor a transition from reactive to proactive oncologic care.</p>

Explainable machine learning framework integrating PPPM for automated cytological triage and personalized breast cancer risk stratification

期刊论文
发表日期：

2025年12月4日

查看全文

Fernando Rodrigues Trindade Ferreira,
Loena Marins do Couto

摘要

Delayed recognition of morphometric abnormalities in breast cytology can contribute to disease progression and adverse clinical outcomes. Conventional cytopathological assessment is affected by subjectivity and inter observer variability, which highlights the need for automated and objective diagnostic support. This study proposes an explainable machine learning framework designed to analyze fine needle aspiration (FNA) cytomorphometric data in order to classify breast lesions and stratify malignancy risk. The Breast Cancer dataset was reorganized into four clinically relevant categories: Typical Benign, Atypical Benign, Early Malignant, and Advanced Malignant. This classification was defined based on hierarchical clustering applied to 30 nuclear and structural morphometric features. Class imbalance was addressed using SMOTE, Gaussian noise augmentation, and Mixup. Five lightweight classifiers were evaluated: LightGBM, XGBoost, multilayer perceptron, random forest, and AdaBoost. Hyperparameters were optimized through grid search with nested cross validation. Model performance was assessed using accuracy, precision, recall, F1 score, and area under the ROC curve (AUC), evaluated both on an independent test set and through stratified five fold cross validation. Statistical significance in AUC differences was examined using the DeLong test. Interpretability was ensured through SHAP, which provided global and instance level explanatory insights. LightGBM presented the best overall performance, achieving a mean accuracy of 97.07% (±0.62), an F1 score of 95.41% (±1.24), and an AUC of 0.987 (±0.004). Its advantage over XGBoost was statistically significant, with an AUC difference of 0.024 and \(p = 0.017\) . SHAP analysis indicated that concavity_mean, area_worst, and perimeter_worst were the most influential morphometric features associated with malignant lesions. The proposed approach introduces a transparent and efficient clinical decision support system for automated triage of breast lesions based on FNA derived morphometric data. The model enhances diagnostic precision and enables malignancy risk stratification, aligned with the principles of Predictive, Preventive and Personalized Medicine (PPPM/3PM). Its explainable design, computational efficiency, and compatibility with telemedicine and digital cytology workflows support its integration into routine practice and favor a transition from reactive to proactive oncologic care.